Evolutionary Robotics and Neuroscience

When research in evolutionary robotics (ER) initially took off in the early 1990s, concerns over the brittleness of traditional AI techniques had recently led to a resurgence of interest in artificial neural networks (ANNs). This fact, coupled with the obvious (loose) analogy between robot control systems and biological nervous systems, meant that most ER researchers naturally gravitated towards neurocontrol systems (Husbands and Harvey 1992, Beer and Gallagher 1992, Harvey et al. 1994, Parisi and Nolfi 1993, Floreano and Mondada 1994). To many of those researchers neural networks also intuitively seemed to be more evolvable than other possible control substrates such as rules or programs – nodes and connections could be gradually changed or added or deleted in a flexible open-ended way (Harvey et al. 1993, Cliff et al. 1993, Beer and Gallagher 1992). In addition, from the earliest days, it has been noted that dynamical recurrent varieties of neural networks, many strongly biologically influenced, allow subtle dynamics that can be readily exploited in the generation of robust adaptive behaviour (de Garis 1990, Beer and Gallagher 1992, Harvey et al. 1993). Hence, from the outset artificial neural networks have been the predominant control system used in ER. Therefore the field has always had at least a tentative link with neuroscience. However, strands of work in which the link is more explicit have existed since the inception of the field and have continued to develop. They are the focus of this chapter.